Conference PaperPDF Available

Data-Driven Acoustic Design of Diffuse Soundfields: Self-Organizing Maps as an exploratory design tool for big data

Authors:

Abstract

The paper demonstrates a novel approach to performance-driven acoustic design of architectural diffusive surfaces. It uses unsupervised machine learning techniques to analyze and explore the GIR Dataset, an extensive collection of real impulse responses and acoustically diffusive surfaces. The presented approach enables designers to explore many alternative acoustically-informed material patterns with various diffusive properties without requiring expert knowledge in acoustics. The paper introduces the computational pipeline, describes the used methods and presents two use-cases in the form of design experiments. Finally, the paper discusses the challenges of developing such a method, its advantages, limitations, and future work.
170 ACADIA 2021
2ACADIA 2021
Data-Driven Acoustic Design
of Diffuse Sound elds
1 3D visualization of a portion
of GIR Dataset's diffusive
surfaces demonstrating possible
geometric variations. For scale
reference, a human is placed at
the bottom left part of the image.
1
ABSTRACT
The paper demonstrates a novel approach to performance-driven acoustic design of
architectural diffusive surfaces. It uses unsupervised machine learning techniques to
analyze and explore the GIR Dataset, an extensive collection of real impulse responses
and acoustically diffusive surfaces. The presented approach enables designers to explore
many alternative acoustically-informed material patterns with various diffusive properties
without requiring expert knowledge in acoustics. The paper introduces the computational
pipeline, describes the used methods, and presents two use-cases in the form of design
experiments. Finally, the paper discusses the challenges of developing such a method, its
advantages, limitations, and future work.
Achilleas Xydis
ETH Zurich
Nathanaël Perraudin
Swiss Data Science Center
Romana Rust
ETH Zurich
Beverly Ann Lytle
ETH Zurich
Fabio Gramazio
ETH Zurich
Matthias Kohler
ETH Zurich
Self-Organizing Maps as an Exploratory Design Tool for Big Data
DATA, BIAS & ETHICS 171
REALIGNMENTS
3
REALIGNMENTS
DATA, BIAS & ETHICS
INTRODUCTION
During the design phase, architects examine a wide range
of alternative design ideas. Early-stage design decisions
signicantly impact the nal design’s performance,
whereas late-stage design modications can rarely
compensate for poor early-stage choices. In fundamental
building components such as structure or façade, perfor-
mance is an integral design driver, which is included
early on. Usually, this is done in close collaboration with
experts in an iterative process where a design is analyzed,
evaluated, and adjusted to meet the desired performance
criteria. This process has become a standard practice
for most architectural projects because architects are
trained to understand these topics (e.g., structural design).
However, room acoustics are rarely included in the early
design process, even though they signicantly impact our
perception of space and well-being. Apart from cases
where sound quality is critical (e.g., concert halls, audi-
toriums), acoustics are either not included in the design
process or come as an afterthought relying on standard-
ized solutions in the form of absorbent or diffusive panels.
The acoustic quality of a room is determined by its geometry
and the structure or pattern of its surfaces. Slight manipu-
lations of the surface geometry could yield signicant gains
in acoustic quality (Cox and D’Antonio 2004). Currently,
through computational design and digital fabrication,
architects already design, visualize, and fabricate surfaces
with complex geometries. While these geometries are
not designed with acoustics in mind, they could act as
sound diffusers, enhancing the room’s acoustic qualities.
Diffusive surfaces reect sound in multiple directions
and, by doing so, reduce echoes, standing waves, and
sound coloration while promoting spaciousness. Suppose
acoustics were included as a design criterion. In that
case, these complex geometries could be an integral part
of architectural elements and combine multiple acoustic
properties, targeting the acoustic needs of their immediate
surrounding.
To employ acoustic performance as a design driver, we
must be able to quantify and interpret the acoustic
effects of our geometric design choices. In a classical
design process, architects have no starting point for an
acoustically performative design of surfaces as they lack
expert knowledge. Different computer simulation software
(Odeon,1 Pachyderm,2 CATT-acoustics3) can be used to
analyze and characterize the acoustic performance of
digitally designed geometries. Nevertheless, this paradigm
relies on the premise that the user is knowledgeable in
room acoustics and knows what adjustments need to be
made to achieve the desired goal. As a result, architects
are discouraged from using such software to evaluate
their design, especially early on. Furthermore, no CAD
nor acoustic simulation software exists that proposes a
geometrical solution to an acoustical question. Further
effort is needed to increase acoustic performance aware-
ness in architecture and provide architects with simple and
accessible workows for designing diffusive surfaces.
Machine Learning (ML) has enabled signicant break-
throughs in automated data processing and pattern
recognition within various elds (Voulodimos et al. 2018;
Arulkumaran et al. 2017; Niu et al. 2019). Architecture and
engineering have also seen an increase in research on how
to employ ML techniques in performance-based design
(Tamke, Nicholas, and Zwierzycki 2018), style transferring
(Steinfeld 2017; Gatys, Ecker, and Bethge 2015), and
clustering (Saldana Ochoa et al. 2020). In ML techniques,
the quality and size of the dataset heavily inuence the
ML model’s nal quality (Mirvis 2002). Although a larger
dataset is desirable—as it makes for a more condent
prediction—the larger the dataset is, the more challenging
it is to navigate, especially for non-expert users. Given
its success in other elds, ML is also used in acoustics
research, mainly as a predictive tool. Datasets built for
this purpose could also be explored as a knowledge base
of known acoustic properties. This paper combines data
clustering techniques with a large dataset of geometries
and impulse responses. It provides an exploratory design
tool for diffusive surfaces, bringing acoustic perfor-
mance-based evaluation earlier into the design stage.
BACKGROUND
Acoustics
In recent years, signicant research has been carried
out on acoustic performance-based design. Shtrepi et al.
(2020) presented a design process that provides architects
and designers with rapid visual feedback on the acoustic
performance of diffusive surfaces. Peters (B. Peters 2015;
Isak Worre Foged 2014) demonstrated methods that
allow tuning acoustic performance while geometry and
materials change. Badino et al. (Badino, Shtrepi, and Astol
2020) presented the state-of-the-art of acoustic perfor-
mance-based design application in practice using nineteen
built projects. Most of these projects were conducted by big
architectural rms in collaboration with expert acoustic
consulting groups but were only geared towards spaces
intended for music performance. Several computational
tools exist that enable the design and optimization of acous-
tically diffusive surfaces. However, their primary focus
is phase grating surfaces (stepped diffusers, quadratic
residue diffusers, primitive root diffusers) (Cox and
D’Antonio 2000), based on sound diffusers introduced by
172 ACADIA 2021
4ACADIA 2021 Data-Driven Acoustic Design of Diffuse Soundelds Xydis, Perraudin,
Rust, Lytle, Gramazio, Kohler
2
Schroeder (Schroeder 1975). Although these tools simplify
the acoustic design process, the generated diffusers have
particular and limited geometries, a substantial thickness,
and a dedicated placement according to acoustic criteria.
These factors make them unattractive and difcult to
integrate into an architectural design that is not purely
focused on music performance.
Machine Learning
The main ML applications in room acoustics have focused
on characterization, information extraction, or classi-
cation. For example, ML has been used to extract the
Reverberation Time and the Early Decay Time of a room
from music signals (Gamper and Tashev 2018) and the
room volume (Genovese et al. 2019). Peters et al. and
Papayiannis et al. (N. Peters, Lei, and Friedland 2012;
Papayiannis, Evers, and Naylor 2020) presented methods
to identify the room type from an audio recording.
Most of the contributions above used supervised learning,
which generally requires large amounts of labeled data.
Data such as impulse response, absorption and scattering
coefcient, early decay time, and many more are primarily
quantitative in nature, therefore, hard to evaluate by
non-acousticians. Moreover, architectural design is often
focused on qualitative measures that depend on the appli-
cation context and the designer’s personal preferences.
Alternately, unsupervised learning allows the extraction of
information from data even when no labels are available.
For example, dimensionality reduction organizes high-
dimensional data samples in a lower-dimensional space—
also known as embedding—by clustering similar samples
together. A high-dimensional space contains data samples
with multiple attributes; for example, an image with a
resolution of just 100 by 100 pixels is a 10000-dimen-
sional sample if we view each pixel within the image as an
attribute. Classical dimensionality-reduction techniques
include Principal Component Analysis (Wold, Esbensen,
and Geladi 1987), t-SNE (Van der Maaten and Hinton 2008),
or Self-Organizing Maps (SOM) (Kohonen 1982). SOMs
have been successfully employed in several elds such as
environmental studies (Gorgoglione et al. 2021), cancer
research (Mazin et al. 2021), chemistry (Motevalli, Sun,
and Barnard 2020), structural design (Saldana Ochoa et al.
2020), and architectural design (Kobayashi 2006). SOMs
are particularly useful in this context. They use unsuper-
vised training to create a nonlinear data transformation
of a high-dimensional space to a low-dimensional space
(usually a two-dimensional map) while preserving the
topological relationships of the original high-dimensional
space (Moosavi 2017). Topology preservation implies that
if two data points are close in the high-dimensional space,
they must also be near each other in the new low-dimen-
sional space and, therefore, belong to the same cluster.
This reduction in complexity makes it possible for designers
to associate a qualitative measure on the embedding.
Dataset
As mentioned in the introduction, the success of ML tech-
niques relies heavily on the quality and size of the dataset
they use. Several acoustic datasets exist containing room
impulse responses (IRs), but their main application is in
the eld of speech enhancement and speech recognition:
AIR,4 BUT ReverdDB,5 RWCP6 (Jeub, Schafer, and Vary
2009; Szoke et al. 2019; Nakamura et al. 2000); acoustic
environment characterization, ACE Corpus7 (Eaton et al.
2017); or for smart-home applications, DIRHA8 (Ravanelli
et al. 2015). Furthermore, these datasets do not contain
any three-dimensional geometrical data. The open-sourced
GIR Dataset9 (Xydis et al. 2021), an extensive collection
of three-dimensional diffuse surfaces and their corre-
sponding real impulse responses, was recently released.
It can be used for ML applications to predict the acoustic
properties of three-dimensional surfaces.
METHODS
As highlighted in sections 1 and 2, acoustic performance
criteria are mainly considered in projects where spaces
host music performances. Furthermore, current methods
mainly focus on design optimization and heavily rely on
expert knowledge in acoustics. This research presents
DATA, BIAS & ETHICS 173
REALIGNMENTS
5
REALIGNMENTS
DATA, BIAS & ETHICS
a workow that enables architects to explore several
possible design solutions, given specic acoustic perfor-
mance criteria (energy per frequency band). It uses the GIR
Dataset for its unique set of three-dimensional surfaces
and the high number of real IRs per surface. Machine
learning techniques and specically SOMs are used to
cluster the surfaces based on acoustic performance
criteria.
The GIR Dataset
The GIR Dataset contains 873,496 real impulse responses
from 296 surfaces (2951 per surface), spread in three
layers (Figure 2). Layer_0 contains 36 measurements in
a 6x6 grid and is the closest to the surface at a distance
of 1 meter. Layer_1 and layer_2 contain 25 and 16
measurements in a 5x5 and 4x4 grid at a distance of
1.9 meters and 2.8 meters, respectively. The IRs were
captured inside a semi-anechoic room and time-windowed
only to contain the rst reections. The surfaces of the
dataset resemble architectural material systems and are
arranged in nine typologies such as brick walls, stone
walls, and more (Figure 3). The geometry of each surface
is composed of a microstructure and a macrostructure.
The rst denes the typology and the second its overall
shape. Several typology-specic material and construction
characteristics are coded in the geometry generation
algorithm and used to create different material patterns.
The brick dimensions, its rotation along the z-axis, its shift
along the macrostructure’s normal vector, and the width
and depth of the mortar are used for the brick typologies.
3
2 The measuring grid in front of a
surface. Red represents the source
position and blue the selected
receivers’ layer. The source is
located in the center of the surface
and 4.5 meters away from it. The
receivers’ layer (layer_1) is approx-
imately 1.9 meters away from the
surface.
3 A sample of different surface
typologies. From top left to bottom
right: Polygonal rubble stones, PRD
diffuser, IDL, Stretcher bond bricks,
Coursed ashlar stones, Primitives,
IDL, Flemish bond bricks.
4 Micro-macrostructure. Left: A
surface with only a microstructure
(Stretcher bond bricks). Middle: A
surface with only a macrostructure.
Right: A surface that combines the
micro- and macrostructure.
4
174 ACADIA 2021
6ACADIA 2021
The number of stones per square meter is used for the
stone typologies, along with the surface roughness and the
joint depth between them. The macrostructure enhances
the low-frequency diffusion by signicantly increasing the
depth variation (Figure 4).
PROPOSED DESIGN WORKFLOW
The proposed design workow contains three main steps:
data preparation, clustering, and design exploration
(Figure5). We use the impulse responses of the GIR Dataset
to compute the primary performance criteria for our
method. The large size of the dataset dictates the need
for data reduction strategies. We use the open-sourced
MiniSom Python library (Vettigli 2018) and create several
custom data visualization algorithms for the clustering step.
These algorithms provide an easy and understandable way
to visualize complex and high-dimensional data and validate
the quality and performance of several steps of our work-
ow. Finally, we describe using the trained SOM to explore
design options based on given acoustic criteria.
Data Preparation
The principal challenge when constructing a low-dimen-
sional embedding using a SOM is the size of a sample.
Given that each of the patterns contains 2951 impulse
responses of 400 oat numbers, the total size of the raw
feature vector is 1,180,400 (Figure 5a). We use three steps
to reduce this large dimension. First, a source position
is selected from the measuring grid, which yields 83
measurements spread across the three grid layers (Figure
5b). In the second step, we use the post-processing pipeline
from Rust et al. (2021) to build low-dimensional feature
vectors from the selected IRs. This pipeline removes the
direct sound from the IR, retaining only the reected sound
coming from the surface. A custom-designed band-pass
lter is used to split the above-mentioned acoustic descrip-
tors into ve frequency bands, with center frequencies
at 250Hz, 500Hz, 1kHz, 2kHz, and 4kHz.10 As a last step,
we use the provided functions to convert the IR into total
normalized cumulative energy11 (TNCE). This step effectively
reduces each IR’s 400 samples to 5 numbers (Figure 5c).
Finally, we concatenate the features of each pattern and
obtain feature vectors of size 36x5, 25x5, 16x5 for layers 0,
1, and 2, respectively (Figure 5d).
Clustering
Clustering operations aim to group various design
options into sets with similar features (in this case, TNCE).
Analogous to the clustering methods used by (Saldana
Ochoa et al. 2020; Fuhrimann et al. 2018), this paper
proposes a method to cluster multiple design options
based on their acoustic performance. Therefore, one can
expect similar acoustic performance for all the designs
of the same cluster. Such clustering can be used as a
data-driven catalog that enables designers to explore the
available design space based on acoustic criteria. The SOM
algorithm organizes all the patterns on a two-dimensional
plane. Figure 6 shows the embedding of 296 patterns based
on TNCE values. As highlighted with the colored outline, the
macrostructure is one of the most discriminative features
for the SOM.
Design Exploration
Using the two-dimensional SOM described in the previous
subsection, designers can get a fast and precise overview
of possible design options. Each cell of the SOM contains
a group of design options clustered based on the acoustic
performance feature selected by the designer (e.g., TNCE).
The hypothetical examples described below are used to
illustrate the proposed design workow.
5 Data preparation pipeline
Data-Driven Acoustic Design of Diffuse Soundelds Xydis, Perraudin,
Rust, Lytle, Gramazio, Kohler
DATA, BIAS & ETHICS 175
REALIGNMENTS
7
REALIGNMENTS
DATA, BIAS & ETHICS
We imagine a generic meeting room where one of its walls
may be freely designed to improve the room’s acoustical
properties. For our performance criterion, we choose
the TNCE values of layer_1 because they are located very
close to the center of the room. Because the IRs contain
only early reection information (see the GIR Data Set),
the TNCE values also contain only the energy from these
early reections. Although the form of the room does not
inuence our method, for simplicity, the meeting room has
a shoebox shape measuring 5 meters wide, 6 meters long,
and 4 meters high. We consider the reected energy of a
at surface as our 100 percent reference (maximum spec-
ular reection). The criterion is the reected energy of the
desired surface, represented as a ratio of the at surface’s
energy. Values higher than 100 percent represent ampli-
cation and lower values energy reduction.
Scenario A does not have a specic material system in
mind, but scenario B assumes designers have already
decided on a material system, specically, a brick wall.
These different decisions result in two different sets of
panels for the SOM training. Scenario A uses all the dataset
typologies, resulting in 279 surfaces, and scenario B only
the brick wall typologies, resulting in 146 surfaces. For the
SOM training, the MiniSom library requires us to provide
values for the following arguments: map dimensions (x, y
number of neurons), training iterations, the neighborhood
function, the sigma, and the learning rate12 (Table 1). Sigma
denes the spread of the neighborhood function in number
of neighbors. The appropriate value for sigma varies by
map dimensions. When the sigma value is too small, the
samples cluster near the center of the map; when it is too
large, the map exhibits several large empty areas towards
the center (Hearty 2016). The learning rate denes only the
initial value of the learning rate for the SOM. With every
training iteration, the learning rate adjusts according to the
following function:
learning rate(t) = learning rate / (1 + t / (0.5 x iterations))
We iterated over different training values to achieve an
optimum embedding (Figures 7 and 8). A SOM with many
neurons has enough space to arrange the data samples.
When multiple very similar samples exist, the SOM
neurons iterations neighborhood
function
sigma learning rate training time
x y
scenario A 10 10 100,000 Gaussian 0.8 1.5 28 sec
scenario B 7 7 100,000 Gaussian 1.0 2.5 22 sec
Table 1: SOM training values
6 An SOM of 296 surfaces based on the TNCE values of layer_1. The displayed surfaces are colored from violet to orange to represent their depth
and the colored outline indicates their macrostructure.
176 ACADIA 2021
8ACADIA 2021
8 3D Visualization of option 1 from scenario A (panel_0100_0) 9 3D Visualization of option 3 from scenario A (panel_0082_0)
7 Left: The 10x10 SOM for Scenario A. The black outline indicate the best matching cell acording to the desired energy values.
Right-top: The surfaces of the best matching cell. Right-bottom: Energy ratios of all matching surfaces.
algorithm places these samples in the same cell; thus, the
resulting embedding can have several empty cells. On the
other hand, a SOM with a very small number of neurons
may not have enough space to arrange the samples. This
constraint will force the algorithm to place less similar
samples on the same cell, resulting in a less representative
data embedding.
Scenario A
This scenario aims to design a surface that, compared to
a at reective surface, lowers the specularly reected
energy in the whole spectrum, emphasizing the mid- and
high-frequency bands. This emphasis will make the room
sound softer by reducing the often harsh high-frequency
specular reections. Combined with the overall reduction
in reected energy, the person speaking will sound more
clear. To achieve the desired energy goal, we input the
following values: [80, 80, 70, 60, 60], and the SOM cell
with the closest matching values is displayed (Figure 7).
Selecting the cell reveals all the surfaces with similar
values in descending order, from the closest to the least
matching option. Nevertheless, because of how the SOM
clustering algorithm works, even the least matching option
is very close to our desired acoustic criterion. Figure 7
shows the energy ratios of all matching surfaces compared
to the desired energies and their close-up views. Option
1 (panel_0100_0) and option 3 (panel_0082_0) are also
visualized inside the room to evaluate them based on
aesthetic qualities. At this point, the architect decides which
surface best suits their design idea.
Data-Driven Acoustic Design of Diffuse Soundelds Xydis, Perraudin,
Rust, Lytle, Gramazio, Kohler
DATA, BIAS & ETHICS 177
REALIGNMENTS
9
REALIGNMENTS
DATA, BIAS & ETHICS
11 3D visualization of option 1 from scenario B (panel_0104_1) 12 3D Visualization of option 2 from scenario B (panel_0036_1)
10 Left: The 7x7 SOM for Scenario A. The black outline indicates the best matching cell acording to the desired energy values.
Right-top: The surfaces of the best matching cell. Right-bottom: Energy ratios of all matching surfaces.
Scenario B
Like the previous scenario, the performance criterion is
again the TNCE values of layer_1. Figure 10 shows the cell
with the best matching values, the close-up views of the
associated surfaces and their energy ratios. In this case,
the SOM cell contains only three surfaces. Contrary to
scenario A, these surfaces happen to have a macrostruc-
ture, making them more spatially expressive. Options 1 and
3 are from the same typology and have very similar designs
and energy values; therefore, we focus on options 1 and
2. Panel_0036_1 lowers the energy by ve to ten percent
more than the desired energy goal in all frequency bands.
Although panel_0104_1 also lowers the energy a little
more than the set goal in the two lowest frequency bands
(250Hz, 500Hz), it matches the desired goal in the 1kHz and
4kHz frequency bands (see Figure 10). Therefore, option 1
better matches our desired acoustic performance criteria.
RESULTS AND DISCUSSION
We have proposed a novel and fast workow for a
performance-driven acoustic design of diffusive
surfaces. We described its components and how each
of them contributes to the entire workow. We have
demonstrated its application with two design experiments.
These experiments showed that thanks to its visual and
intuitive implementation, users need little acoustic expert
knowledge to specify and explore early design options
compared to traditional room acoustic surface design
processes. When no predened typology is chosen, the
design proposals could include several different typologies.
178 ACADIA 2021
10 ACADIA 2021
This approach could inspire or drive the designer’s choices
and could also be used as a basis for discussion and
further renement with acoustic experts. Compared to
sometimes days of computing time when using numerical
modeling algorithms (B. Peters 2015), our method needs
only 20 to 30 seconds13 to train the SOM depending on
the dataset size (See Table 1). Then, computing the closest
matching designs requires less than a second. Although
the presented workow is based on the GIR Dataset and
a panel’s precise TNCE values, one could use any other
acoustical descriptor from the GIR Dataset (impulse
response, frequency response, cumulative energy, and
more). Furthermore, the presented methodology is not
limited to the GIR Dataset. It can be adapted and applied
to any other acoustic dataset..
The two design scenarios have shown that both at-like
(scenario A) and spatially varied surfaces (scenario B)
are considered options. Flat-like surfaces are more likely
to have uniform TNCE values across points of the same
layer, and are more likely, therefore, to be the closest
matching sample in the SOM. Nevertheless, the design
workow is not limited to a single set of desired energy
values. We can assign different values to each layer, assign
individual values to each grid point of a specic layer, and
nally, assign a few values at desired locations and let the
algorithm interpolate the in-between values. The fewer sets
of energy values one uses as a performance criterion, the
more likely it will result in a at-like design.
Limitations
Although the presented design workow proposes
material patterns based on desired acoustic performance
criteria, these patterns can only be from the GIR dataset.
Nevertheless, the dataset can be expanded to include more
patterns for a specic typology or introduce an entirely
new typology. Furthermore, because the measurements
were not according to the ISO standard, they cannot be
used to derive standard acoustical descriptors such as
absorption and scattering coefcients. Therefore, the
clustering can only be done using the descriptors provided
by Rust et al. (2021) (e.g., cumulative energy, normalized
cumulative energy, tonal normalized cumulative energy).
Nevertheless, we believe that total energy values (TNCE),
split into ve lter bands, are metrics most users can
understand or quickly get familiar with.
Future Work
The proposed design workow provides initial ideas or
inspiration for a more acoustically informed design direc-
tion. However, choosing the desired acoustical parameters
for the different frequency bands may still require some
basic understanding of acoustics or initial consultation
with an acoustics expert. Therefore, predened acoustic
use-cases should be implemented. These cases will trans-
late qualitative intentions into quantitative parameters.
Currently, the design workow can be used via a Jupyter
notebook, and it is available as an open-source code in
https://renkulab.io/gitlab/ddad/ddad-renku/. The interface
can be further streamlined and possibly integrated as a tool
within existing CAD software or a stand-alone web-based
application.
AUTHORS CONTRIBUTIONS
Conceptualization, A.X.; methodology, A.X.; coding A.X., N.P.; code
optimization, B.L.; writing-original draft, A.X.; writing-review and
editing, A.X., R.R., F.G., M.K.; gures and visualization, A.X.;
supervision, R.R., F.G., M.K.
ACKNOWLEDGMENTS
This research stemmed out of a collaborative and multidisciplinary
project between Gramazio Kohler Research at ETH Zurich, the
Swiss Data Science Center, the Laboratory for Acoustics/Noise
Control at EMPA, and STRAUSS ELEKTROAKUSTIK GMBH. Therefore,
the authors would like to thank Dr. Fernando Perez-Cruz, Dr. Kurt
Heutschi, Kurt Eggenschwiler, and Jurgen Strauss for their inputs.
Furthermore, we would like to thank Dr. Nikola Marinčić for his
valuable inputs on self-organizing maps and Gonzalo Casas for
always being keen (hopefully) on helping on Python related topics.
NOTES
All website were accessed on October 8th, 2021.
1. http://www.odeon.dk.
2. http://www.orase.org.
3. http://www.catt.se.
4. https://www.iks.rwth-aachen.de/forschung/tools-downloads/
databases/aachen-impulse-response-database/
5. https://speech.t.vutbr.cz/software/
but-speech-t-reverb-database.
6. http://research.nii.ac.jp/src/en/RWCP-SSD.html.
7. https://acecorpus.ee.ic.ac.uk/.
8. http://dirha.fbk.eu/English-PHdev.
9. https://renkulab.io/projects/ddad/gir-dataset/.
10. The geometries and frequencies in the data set are in 1:10 scale.
11. The TNCE is the last value from the NCE list, representing the
total energy arrived at the receiver position.
12. Further documentation can be found on MiniSom’s Github
repository. https://github.com/JustGlowing/minisom.
13. On a 2.9GHz 6-core Intel i9 cpu and 32GB 2400MHz DDR4 RAM.
Data-Driven Acoustic Design of Diffuse Soundelds Xydis, Perraudin,
Rust, Lytle, Gramazio, Kohler
DATA, BIAS & ETHICS 179
REALIGNMENTS
11
REALIGNMENTS
DATA, BIAS & ETHICS
REFERENCES
Arulkumaran, Kai, Marc Peter Deisenroth, Miles Brundage, and
Anil Anthony Bharath. 2017. “Deep Reinforcement Learning: A Brief
Survey.IEEE Signal Processing Magazine 34 (6): 26–38.
Badino, Elena, Louena Shtrepi, and Arianna Astol. 2020. “Acoustic
Performance-Based Design: A Brief Overview of the Opportunities
and Limits in Current Practice.Acoustics 2 (2): 246–78.
Cox, Trevor J., and Peter D’Antonio. 2000. “Acoustic Phase Gratings
for Reduced Specular Reection.Applied Acoustics 60 (2): 167–86.
Cox, Trevor J., and Peter D’Antonio. 2004. Acoustic Absorbers and
Diffusers: Theory, Design and Application. Boca Raton: CRC Press.
Eaton, James, Nikolay D. Gaubitch, Alastair H. Moore, and Patrick
A. Naylor. 2017. “Acoustic Characterization of Environments (ACE)
Challenge Results Technical Report.ArXiv:1606.03365.
Fuhrimann, Lukas, Vahid Moosavi, Patrick Ole Ohlbrock, and
Pierluigi D’acunto. 2018. “Data-Driven Design: Exploring New
Structural Forms Using Machine Learning and Graphic Statics.
Proceedings of IASS Annual Symposia 2018 (2): 1–8.
Gamper, Hannes, and Ivan J. Tashev. 2018. “Blind Reverberation
Time Estimation Using a Convolutional Neural Network.” In
Proceedings of the 16th International Workshop on Acoustic Signal
Enhancement (IWAENC). Tokyo, Japan. 136–140.
Gatys, Leon A., Alexander S. Ecker, and Matthias Bethge. 2015. “A
Neural Algorithm of Artistic Style.ArXiv:1508.06576.
Genovese, Andrea F., Hannes Gamper, Ville Pulkki, Nikunj
Raghuvanshi, and Ivan J. Tashev. 2019. “Blind Room Volume
Estimation from Single-Channel Noisy Speech.” In ICASSP 2019;
2019 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP). Brighton, UK. 231–35.
Gorgoglione, Angela, Alberto Castro, Vito Iacobellis, and Andrea
Gioia. 2021. “A Comparison of Linear and Non-Linear Machine
Learning Techniques (PCA and SOM) for Characterizing Urban
Nutrient Runoff.Sustainability 13 (4): 2054.
Hearty, John. 2016. Advanced Machine Learning with Python.
Birmingham: Packt Publishing Ltd.
Isak Worre Foged, Anke Pasold, and Mads Brath Jensen. 2014.
“Evolution of an Instrumental Architecture.” In Fusion; Proceedings
of the 32nd ECAADe Conference, vol 2., edited by Emine M.
Thompson. Northumbria University, Department of Architecture
and Built Environment. Newcastle upon Tyne, UK: ECAADe.
365–372.
Jeub, Marco, Magnus Schafer, and Peter Vary. 2009. “A
Binaural Room Impulse Response Database for the Evaluation
of Dereverberation Algorithms.” In 2009 16th International
Conference on Digital Signal Processing. Santorini, Greece. 1–8.
Kobayashi, Yosihiro. 2006. “Self-Organizing Map and Axial Spatial
Arrangement: Topological Mapping of Alternative Designs.” In
Synthetic Landscapes; Proceedings of the 25th Annual Conference
of the Association for Computer-Aided Design in Architecture
(ACADIA). 342–355.
Kohonen, Teuvo. 1982. “Self-Organized Formation of Topologically
Correct Feature Maps.Biological Cybernetics 43 (1): 59–69.
Mazin, Asim, Samuel H. Hawkins, Olya Stringeld, Jasreman Dhillon,
Brandon J. Manley, Daniel K. Jeong, and Natarajan Raghunand.
2021. “Identication of Sarcomatoid Differentiation in Renal
Cell Carcinoma by Machine Learning on Multiparametric MRI.
Scientic Reports 11 (1): 3785.
Mirvis, Stuart E. 2002. “Garbage in, Garbage out (How Purportedly
Great ML Models Can Be Screwed up by Bad Data).Applied
Radiology 31 (2): 5.
Moosavi, Vahid. 2017. “Contextual Mapping: Visualization of High-
Dimensional Spatial Patterns in a Single Geo-Map.Computers,
Environment and Urban Systems 61 (January): 1–12.
Motevalli, Benyamin, Baichuan Sun, and Amanda S. Barnard. 2020.
“Understanding and Predicting the Cause of Defects in Graphene
Oxide Nanostructures Using Machine Learning.The Journal of
Physical Chemistry C 124 (13): 7404–13.
Nakamura, Satoshi, Kazuo Hiyane, Futoshi Asano, Takanobu
Nishiura, and Takeshi Yamada. 2000. “Acoustical Sound Database
in Real Environments for Sound Scene Understanding and
Hands-Free Speech Recognition.” In Proceedings of the Second
International Conference on Language Resources and Evaluation
(LREC’00). Athens, Greece: European Language Resources
Association (ELRA).
180 ACADIA 2021
12 ACADIA 2021
Niu, Haiqiang, Zaixiao Gong, Emma Ozanich, Peter Gerstoft, Haibin
Wang, and Zhenglin Li. 2019. “Deep-Learning Source Localization
Using Multi-Frequency Magnitude-Only Data.The Journal of the
Acoustical Society of America 146 (1): 211–22.
Papayiannis, Constantinos, Christine Evers, and Patrick A. Naylor.
2020. “End-to-End Classication of Reverberant Rooms Using
DNNs.IEEE/ACM Transactions on Audio, Speech, and Language
Processing 28: 3010–17.
Peters, Brady. 2015. “Integrating Acoustic Simulation in
Architectural Design Workows: The FabPod Meeting Room
Prototype.Simulation 91 (9): 787–808.
Peters, Nils, Howard Lei, and Gerald Friedland. 2012. “Name That
Room: Room Identication Using Acoustic Features in a Recording.
In MM ’12: Proceedings of the 20th ACM International Conference
on Multimedia. New York: ACM. 841–44.
Ravanelli, Mirco, Luca Cristoforetti, Roberto Gretter, Marco
Pellin, Alessandro Sosi, and Maurizio Omologo. 2015. “The DIRHA-
ENGLISH Corpus and Related Tasks for Distant-Speech Recognition
in Domestic Environments.” In 2015 IEEE Workshop on Automatic
Speech Recognition and Understanding (ASRU). 275–82.
Rust, Romana, Achilleas Xydis, Kurt Heutschi, Nathanael
Perraudin, Gonzalo Casas, Chaoyu Du, Jürgen Strauss, et
al. 2021. “A Data Acquisition Setup for Data Driven Acoustic
Design.Building Acoustics 28 (4): 345–360. https://doi.
org/10.1177/1351010X20986901.
Saldana Ochoa, Karla, Patrick Ole Ohlbrock, Pierluigi D’Acunto,
and Vahid Moosavi. 2020. “Beyond Typologies, Beyond
Optimization: Exploring Novel Structural Forms at the Interface
of Human and Machine Intelligence.International Journal
of Architectural Computing 19 (3): 466–490. https://doi.
org/10.1177/1478077120943062.
Schroeder, Manfred R. 1975. “Diffuse Sound Reection by
Maximum-length Sequences.Journal of the Acoustical Society
of America 57 (1): 149.
Shtrepi, Louena, Tomás Mendéz Echenagucia, Elena
Badino, and Arianna Astol. 2020. “A Performance-Based
Optimization Approach for Diffusive Surface Topology
Design.Building Acoustics 28 (3): 231–247. https://doi.
org/10.1177/1351010X20967821.
Steinfeld, Kyle. 2017. “Dreams May Come.” In ACADIA 17:
Disciplines & Disruption; Proceedings of the 37th Annual
Conference of the Association for Computer Aided Design in
Architecture (ACADIA). Cambridge, MA. 590–99.
Szoke, Igor, Miroslav Skacel, Ladislav Mosner, Jakub Paliesek, and
Jan “Honza” Cernocky. 2019. “Building and Evaluation of a Real
Room Impulse Response Dataset.IEEE Journal of Selected Topics
in Signal Processing 13 (4): 863–76.
Tamke, Martin, Paul Nicholas, and Mateusz Zwierzycki. 2018.
“Machine Learning for Architectural Design: Practices and
Infrastructure.International Journal of Architectural Computing
16 (June): 123–43.
Van der Maaten, Laurens, and Geoffrey Hinton. 2008. “Visualizing
Data Using T-SNE.Journal of Machine Learning Research 9 (11):
2579–2605.
Vettigli, Giuseppe. 2018. MiniSom: Minimalistic and NumPy-Based
Implementation of the Self Organizing Map. Accessed January 15,
2022. https://github.com/JustGlowing/minisom.
Voulodimos, Athanasios, Nikolaos Doulamis, Anastasios
Doulamis, and Eftychios Protopapadakis. 2018. “Deep
Learning for Computer Vision: A Brief Review.Computational
Intelligence and Neuroscience 2018 (February). https://doi.
org/10.1155/2018/7068349.
Wold, Svante, Kim Esbensen, and Paul Geladi. 1987. “Principal
Component Analysis.Chemometrics and Intelligent Laboratory
Systems, Proceedings of the Multivariate Statistical Workshop for
Geologists and Geochemists 2 (1): 37–52.
Xydis, Achilleas, Nathanaël Perraudin, Romana Rust, Kurt Heutschi,
Gonzalo Casas, Oksana Riba Grognuz, Kurt Eggenschwiler,
Matthias Kohler, and Fernando Perez-Cruz. 2021. GIR Dataset: A
Geometry and Real Impulse Response Dataset [Data set]. Zenodo.
https://doi.org/10.5281/zenodo.5288744.
Data-Driven Acoustic Design of Diffuse Soundelds Xydis, Perraudin,
Rust, Lytle, Gramazio, Kohler
DATA, BIAS & ETHICS 181
REALIGNMENTS
13
REALIGNMENTS
DATA, BIAS & ETHICS
IMAGE CREDITS
All drawings and images by the authors.
Achilleas Xydis is an architect and current a doctoral researcher
at the Chair of Architecture and Digital Fabrication (Gramazio
Kohler Research) at ETH Zurich. He received his Diploma in
Architecture from the University of Patras in 2010. In 2013
he completed the post-graduate Master of Advanced Studies
in Architecture and Information at ETH Zurich. He focuses on
combining machine learning techniques with architectural
acoustics.
Nathanaël Perraudin nished his Master in electrical engi-
neering at the Ecole Fédérale de Lausanne (EPFL) and worked as
a researcher in the Acoustic Research Institute (ARI) in Vienna.
In 2013, he came back to EPFL for a PhD. He specialized in signal
processing, graph theory, machine learning, data science, and
audio processing. After graduating in 2017, he began working as
a senior data scientist at the Swiss Data Science Center (SDSC),
focusing on deep neural networks and generative models.
Romana Rust is a computational architect and senior researcher
at Gramazio Kohler Research, ETH Zurich within the Design++
initiative: Centre for Augmented Computational Design in AEC.
She is the co-coordinator of the Immersive Design Lab, a lab
for collaborative research and teaching in the eld of extended
reality and machine learning in architecture and construction. Her
particular interest is the development of innovative computational
methods that integrate multiple design objectives such as geom-
etry, acoustics, materiality, and robotic fabrication.
Beverly Ann Lytle is a software engineer at the Chair for
Architecture and Digital Fabrication at ETH Zurich, where she
contributes to COMPAS, an open-source python framework for
AEC. She received her master's degree at Ohio State University
and her PhD at ETH Zurich, both in mathematics.
Fabio Gramazio and Matthias Kohler are professors of
Architecture and Digital Fabrication at ETH Zurich. In 2000, they
founded the architecture practice Gramazio & Kohler, which
realized numerous award-winning projects. Opening the world’s
rst architectural robotic laboratory at ETH Zurich, Gramazio &
Kohler’s research has been formative in the eld of digital architec-
ture, setting precedence and de facto creating a new research eld
merging advanced architectural design and additive fabrication
processes through the customized use of industrial robots.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Sarcomatoid differentiation in RCC (sRCC) is associated with a poor prognosis, necessitating more aggressive management than RCC without sarcomatoid components (nsRCC). Since suspected renal cell carcinoma (RCC) tumors are not routinely biopsied for histologic evaluation, there is a clinical need for a non-invasive method to detect sarcomatoid differentiation pre-operatively. We utilized unsupervised self-organizing map (SOM) and supervised Learning Vector Quantizer (LVQ) machine learning to classify RCC tumors on T2-weighted, non-contrast T1-weighted fat-saturated, contrast-enhanced arterial-phase T1-weighted fat-saturated, and contrast-enhanced venous-phase T1-weighted fat-saturated MRI images. The SOM was trained on 8 nsRCC and 8 sRCC tumors, and used to compute Activation Maps for each training, validation (3 nsRCC and 3 sRCC), and test (5 nsRCC and 5 sRCC) tumor. The LVQ classifier was trained and optimized on Activation Maps from the 22 training and validation cohort tumors, and tested on Activation Maps of the 10 unseen test tumors. In this preliminary study, the SOM-LVQ model achieved a hold-out testing accuracy of 70% in the task of identifying sarcomatoid differentiation in RCC on standard multiparameter MRI (mpMRI) images. We have demonstrated a combined SOM-LVQ machine learning approach that is suitable for analysis of limited mpMRI datasets for the task of differential diagnosis.
Article
Full-text available
In this paper, we present a novel interdisciplinary approach to study the relationship between diffusive surface structures and their acoustic performance. Using computational design, surface structures are iteratively generated and 3D printed at 1:10 model scale. They originate from different fabrication typologies and are designed to have acoustic diffusion and absorption effects. An automated robotic process measures the impulse responses of these surfaces by positioning a microphone and a speaker at multiple locations. The collected data serves two purposes: first, as an exploratory catalogue of different spatio-temporal-acoustic scenarios and second, as data set for predicting the acoustic response of digitally designed surface geometries using machine learning. In this paper, we present the automated data acquisition setup, the data processing and the computational generation of diffusive surface structures. We describe first results of comparative studies of measured surface panels and conclude with steps of future research.
Article
Full-text available
Urban stormwater runoff represents a significant challenge for the practical assessment of diffuse pollution sources on receiving water bodies. Given the high dimensionality of the problem, the main goal of this study was the comparison of linear and non-linear machine learning (ML) methods to characterize urban nutrient runoff from impervious surfaces. In particular, the principal component analysis (PCA) for the linear technique and the self-organizing map (SOM) for the non-linear technique were chosen and compared considering the high number of successful applications in the water quality field. To strengthen this comparison, these techniques were supported by well-known linear and non-linear methods. Those techniques were applied to a complete dataset with precipitation, flow rate, and water quality (sediments and nutrients) records of 577 events gathered for a watershed located in Southern Italy. According to the results, both linear and non-linear techniques can represent build-up and wash-off, the two main processes that characterize urban nutrient runoff. In particular, non-linear methods are able to capture and represent better the rainfall-runoff process and the transport of dissolved nutrients in urban runoff (dilution process). However, their computational time is higher than the linear technique (0.0054 s vs. 15.24 s, for linear and non-linear, respectively, in our study). The outcomes of this study provide significant insights into the application of ML methods for the water quality field.
Article
Full-text available
Current development in digital design, combined with the growing awareness of the importance of building performance, had drawn attention to performance-based design (PBD) in architecture. PBD benefits both design workflow and outcome, allowing one to control the performance of the design proposal since early design phases. The paper aims to explore its current application in the acoustic field, where its potential is still little exploited in architectural practice. A set of built case studies is collected and briefly analyzed with the aim to shed some light on the state of the art of the application of acoustic performance-based design (APBD) in practice. The analysis suggests that in order to encourage the application of APBD it is needed on one side to enhance the integration and interoperability among modeling and simulation tools, and on the other side to improve the acoustic knowledge and programming skills of the architectural practitioners.
Article
Different numerical techniques have been used in the last decades for the acoustic characterization and performance optimization of sound diffusive surfaces. However, these methods require very long calculation times and do not provide a rapid feedback. As a result, these methods can hardly be applied by designers at early stages of the design process, when successive design iterations are necessary from an aesthetic point of view. A suitable alternative could be the use of parametric modeling in combination with performance investigations during the design process of sound diffusive surfaces. To this aim, this study presents a design process for diffusive surfaces topology optimization based on the combination of parametric models and geometrical acoustic simulations. It aims to provide architects and designers with rapid visual feedback on acoustic performances at a preliminary stage of the design process. The method has been tested on different case studies, which have been modelled based on geometric guidelines for diffusive surface optimization. The sensitivity of the method showed that it could be a very useful tool for comparisons among surface design alternatives. Finally, the advantages and limitations of the integrated optimization in comparison with conventional optimizations are discussed.
Article
Reverberation is present in our workplaces, our homes, concert halls and theatres. This paper investigates how deep learning can use the effect of reverberation on speech to classify a recording in terms of the room in which it was recorded. Existing approaches in the literature rely on domain expertise to manually select acoustic parameters as inputs to classifiers. Estimation of these parameters from reverberant speech is adversely affected by estimation errors, impacting the classification accuracy. In order to overcome the limitations of previously proposed methods, this paper shows how DNNs can perform the classification by operating directly on reverberant speech spectra and a CRNN with an attention-mechanism is proposed for the task. The relationship is investigated between the reverberant speech representations learned by the DNN and acoustic parameters. For evaluation, AIRs are used from the ACE-challenge dataset that were measured in 7 real rooms. The classification accuracy of the CRNN classifier in the experiments is 78% when using 5 hours of training data and 90% when using 10 hours.<br/
Article
This article presents a computer-aided design framework for the generation of non-standard structural forms in static equilibrium that takes advantage of the interaction between human and machine intelligence. The design framework relies on the implementation of a series of operations (generation, clustering, evaluation, selection, and regeneration) that allow to create multiple design options and to navigate in the design space according to objective and subjective criteria defined by the human designer. Through the interaction between human and machine intelligence, the machine can learn the nonlinear correlation between the design inputs and the design outputs preferred by the human designer and generate new options by itself. In addition, the machine can provide insights into the structural performance of the generated structural forms. Within the proposed framework, three main algorithms are used: Combinatorial Equilibrium Modeling for generating of structural forms in static equilibrium as design options, Self-Organizing Map for clustering the generated design options, and Gradient-Boosted Trees for classifying the design options. These algorithms are combined with the ability of human designers to evaluate non-quantifiable aspects of the design. To test the proposed framework in a real-world design scenario, the design of a stadium roof is presented as a case study.
Article
Machine learning is a powerful way of uncovering hidden structure/property relationships in nanoscale materials, and it is tempting to assign structural causes to properties based on feature rankings reported by interpretable models. In this study of defective graphene oxide nanoflakes we use classification, regression and causal inference to show that not all important structural features directly influence concentration of broken bonds, as a representative property. We find that while the presence of oxygen is important for actual bond breakage, the presence and distribution of hydrogen determines how often bond breakage occurs.